A STUDY OF NON-PERFORMING LOAN BEHAVIOR IN P2P LENDING UNDER ASYMMETRIC INFORMATION

Transcription

1 A STUDY OF NON-PERFORMING LOAN BEHAVIOR IN P2P LENDING UNDER ASYMMETRIC INFORMATION Zongfeng Zou School of Management, Shanghai University, P.R.China Huixin Chen School of Management, Shanghai University, P.R.China Xiaosong Zheng* SHU-UTS SILC Business School Abstract Recent years have witnessed the popularity of online Peer-to-Peer (P2P) lending which is regarded an alternative to banks where individual members lend and borrow money using an online trading platform without official financial intermediaries involved. The paper studies P2P lending and the determinants of Non-performing Loans (NPLs). Different from traditional loans, borrowers in P2P should bear higher credit risks because of information asymmetry. Besides, NPLs are the direct financial lost for borrowers. It is important to establish a reliable model to examine the determinants of NPLs based on the published information of borrowers in the platform. The empirical study is based on the transaction data of PPDai, a leading P2P platform of China in 2015 through non-parametric random forest (RF) model which has superiority in dealing with complex multidimensional data set and higher predict accuracy. A binary logistic regression model is also generated to quantify the possibility of NPLs based on the selected indicator through RF. According to the result, the times of payment are the most predictive factor to NPLs. Keywords: Information asymmetry; non-performing loans (NPL); P2P lending JEL code: D12 1. Introduction Online peer to peer (P2P) lending has gained wide attention in the past few years. It is known as a special type of microloan without the intermediary like commercial bank (Cai et al., 2015). It has been introduced as a new e-commerce phenomenon in the financial field with the support of an internet-based information system which has the potential to provide more economical efficiencies (Wang et al. 2009).In P2P online platform, borrowers submit application for loans, which contains borrowers personal information and loan information. appeal directly to a pool of individual lenders using a request which can include textual narratives to justify the loan. Normally in a P2P transaction, individuals who would like to

2 borrow money (hereafter referred to as borrowers ) and those who would like to lend money (hereafter referred to as lenders ) have no previous relationship (Larimore et al., 2011). Compared with traditional loans, there are several benefits of P2P lending. From lenders perspective, they can obtain higher return on their investments. Light (2012) reported that some leading P2P platform have provided 10% or higher annual returns to the investors at a time of historically low interest rate. As for borrowers, they can borrow money at a lower interest rate. Besides, the threshold of P2P loan is relatively lower than bank loan. For example, borrowers who cannot get loans from bank can acquire loans from P2P lending platform (Wang et al. 2009). Hence, there are more borrowers whose with poor credibility attracted by P2P lending (Schreiner, 2000). As a result, the lenders would bear higher risks. Moreover, lenders would not know the actual credibility of those borrowers because of information asymmetry. The previous researches The most direct result of credit risk is Non-performing loans (NPLs) which is the actual economic lost of lenders. According to Us (2016), NPL is regarded as an indicator of financial weakness. It is important to reduce the NPL rate to ensure financial stability. The current P2P lending platforms in China do not make a connection with China credit information system, they cannot screen out any potential high-risk borrowers based on the official credit score as Lending Club does (Emekter et al. 2015). It is crucial to find a way to reduce the credit risk before lost happens. The purpose of the paper is to conduct an empirical study to explore the common characteristics of NPL borrowers and establish a model to quantify the possibility of NPL of a borrower based on PPDai, one of the leading P2P lending platform in China. The contributions of the paper to the emerging P2P lending market of China in many ways. Firstly, the previous researches focus more on the credit score model building and the factors for successful loan. Few studies concern the problem of NPLs, especially the factors to NPL. Secondly, those empirical studies are mainly based on foreign P2P dataset like the Lending Club instead of apply dataset from China market. Thirdly, few of current P2P researches in China quantify the NPL possibility, the majority of them are qualitative. The rest of the paper is organized as follows: section 2 briefly reviews and summarizes some relevant studies on P2P lending and credit risks. Section 3 introduces the data sample used in the paper and also generally introduce the non-parametric random forest method. Section 4 conducts an empirical study based on the data processed in section 3, the result also displayed as well. Finally, section 5 offers practical implications, scholarly contributions, limitations of the study, and future directions. It also concludes the work. 2. Literature review A large amount of researches have emerged in response to the popular P2P lending market.

3 The theoretical basis of the paper is information asymmetry. Gorton and Winton (2003) argued that information asymmetry is an important feature of credit markets. Spence (2001) suggested that informational gaps between borrowers and lenders can result in adverse selection. In traditional loan environment, banks are considered as a significant financial intermediary can effectively reduce information asymmetry because of financial expertise and extensive experiences to evaluated the creditworthiness of borrowers (Diamond, 1984). The problem of information asymmetry is more severe in P2P lending market than in traditional loan market because of non-financial intermediaries involved (Lee and Lee, 2012). How to alleviate information asymmetry is important for P2P lending market s long-term stability. Some scholars suggested the disclosure and the analysis of the borrowers soft information is an effective way to mitigate information asymmetry (Iyer et al. 2009). In order to improve the P2P lending market in China, offline verification mechanism has been already developed and adopted by some leading sites such as PPdai (Qi et al. 2017). Similar to traditional credit market places, risk assessment and decision making can be viewed from the different perspectives (Wu and Hsu, 2012). From the borrower s perspective, a common goal is securing loan funding. Among all the information provided by the borrower, researchers aim to find determinants of the success of a loan. For example, Wen and Wu (2014) conducted an empirical study based on the transaction data of PPDai to examined the factors to successful loan through binary logistic regression model. The result shows that the interest rate of borrowing and the times of borrower s failure borrowing in the past negatively influence successful borrowing rate, while the amount of borrowing, times of borrower s successful borrowing in the past, credit scores, number of censored items positively influence successful borrowing rate. P2P lending is a risky activity for individual lenders, because the loans are granted by them, instead of P2P companies, that is the lenders should bear the credit risk themselves. Credit risk can be defined as the potential financial impact of any real or perceived change in borrowers creditworthiness, while credit worthiness is the borrowers willingness and ability to repay (Lin, Li and Zheng, 2016). There are some of the existing researches that analyze the credit risk and default rate of loans. Strong social networking relationship is an important determinant of successful loan with lower default risk through an analysis of the role of social connection in credit evaluation conducted by Lin et al. (2013). There are some other studies focus on how personal information and loan information can affect the loan status. Oni et al. found that farmers age, income and educational level are major factors to default in loan repayment through an analysis of a random sample if 100 poultry farmers (Oni, Oladele, and Oyewole 2006). Emekter et al. (2014) found that FICO score, credit grade, revolving line utilization and debt-to income ratio play a significant role in loan defaults (Emekter et al. 2014). Xu et al. (2016) conducted an empirical study based on a P2P lending platform in China and proposed a risk evaluation model to quantify credit risk of a borrower. The result shows personal characteristics (such as age, gender, marital status), loan information ( such as monthly payment, loan amount) and delinquency history play an critical role in defaults (Lin, Li and Zheng, 2016)

4 Few researches on NPLs of P2P lending market while the researches in traditional loan market are quite extensive. Klein (2014) suggested that both bank-specific and macroeconomic factors have impacts on NPLs after an analysis of bank-level data of 16 Central, Eastern and South-Eastern European countries. Ghosh (2015) applied both fixed effects and dynamic-gmm estimation to explore both state-level banking-industry specific as well as region economic determinants of NPLs of all commercial banks across 50 US states. The result indicates greater capitalization, liquidity risk can increase NPLs. Tanaka et. al (2016) introduced applied random forest model to develop a novel early warning system for predicting bank failure. The result shows the method they applied outperforms than conventional methods in terms of prediction accuracy. 3. Data analysis The sample data is based on PPDai, a leading P2P platform in China. PPDai has more than 32,610,000 members since funded in The sample contains all of the transaction data from January to December of 2015, which are 114,717 transaction data in total. There are 1,496 non-performing loans in the sample. According to the repayment period and loan time, the in-processing loan and some advanced repayment loan has been eliminated. Finally, there are 8,461 normal loans in the sample as well. Removed the noise and borderline data, the subsample contains 9,276 loans in total, 1,496 non-performing loans and 7,780 normal data. As shown in Table 1, 16.1% of the loans, that is, a total amount of 18.6 million RMB have been lost. According to Table 1, the amount of normal loans is almost five times of NPLs. In order to increase the accuracy of the result, it is necessary to do data balancing to make the ratio of the two classes to be 1:1. The subsample has been divided into training set and testing set randomly. Table 1. Loan distribution by the loan status All loans Number of loans Per cent Amount (yuan) Per cent Non-performing loans 1, ,603, Normal loans 7, ,147, Total 9, ,750, Based on the available data, the variables of each transaction consist of non-performing information, loan information, basic information, credit information, financial information, historical information and certification information. Table 2 displays the specific variables and the data type of each variable.

5 Table 2. Variable sets in the sample Variable set Variable Type Variable set Variable Type Non-performing information Loan information Basic information Credit information loan status binary income categorical amount interest rate numeric numeric Financial information car asset house asset binary binary repayment period numeric car loan binary borrow type categorical house loan binary investment type binary times of application numeric times of age numeric numeric success Historical times of degree categorical information numeric payments total marital status categorical historical numeric loan amount gender binary identity authentication binary working years categorical income certification binary company size categorical Certification working information certification binary credit limit numeric authentication offline binary credit rating categorical credit report binary Non-performing informaiton: only a variable in the dimension, is loan status, which is represented by Y in the analysis process in the paper. It only have two values, 1 represent for NPL and 0 for normal loans. Loan information: there are both numeric and categorical varibles in the dimenison. Amount is the amount of money borrowed by a borrower. Repayment period is the agreed time to repay the loan. Borrow type is actually the purpose of loan, like short-term turnover, education and so on so forth. Investment type is a binary value, 1for individual investment and 0 for not. Basic information: variables in the dimension are some personal characteristics of a borrowers. Age is a numeric type, it will be input into the model directly. While degree is a categorical type, which will be repleced by 1 to 4 respondently. 1 represents for high school or below, 2 for colledge, 3 for bachelor and 4 for master or above. There are 4 marital statuses, 1 for single, 2 for married, 3 for divorced anf 4 for widowed. Gender is a binary type, 1 for

6 male and 0 for female. Working year is a numeric type while company size is a categorical type which has been categorized into 4 scales based on the amount of workers. Credit information: credit limit is the largest amount of money a borrower can borrow from the platform, which is the numeric type. Credit rating is a direct credit status assigned by the platform. Since P2P platform has no access to the official credit system of China, the value is relatively less rational. There are 7 levels of credit rating, 1 for HR represent the lowest level which means the borrow is relatively less trustworthy. 2-7 represent E to AA, while AA is the highest level of credit rating. Credit report is binary. If a borrower has provied the credit report, the value is 1; if not, the value is 0. Financial information: it reflects the general financial situation of a borrower. Income is the numeric type, it would be input into the model directly. Whihle house and car assets and house and car loan are all binary. If a borrower owns a house, the value of house asset would be 1, otherwise if 0. Historical information: all of the four variables in the dimension are numeric. They reflect the historical loan behavior of a borrower, like how many times they have applied and how many times they have repaid the money and so on. Certification information: all of the four variables are binary. If a borrow s income, work and indentify has been certified, the value of each variable is 1, if not, they would be 0. As mentioned before, in order to decrease the impact of information asymmetry PPDai would conduct offline authentication to borrowers. If a borrower has been authenticated offline, the value would be 1, otherwise, 0. In this paper, loan status is represented by Y in the analysis process, and from amount to authentication offline are represented by X1 to X27. As mentioned above, the general NPL rate of this data sample is 16.1%. In order to understand the NPL rate amount different features, the paper conducts a cross analysis between some critical variables and loan status. Those variables are credit rating, repayment period, degree and gender. Although P2P platforms have no access to the official credit system of China, they would assign borrowers with a certain credit rating based on borrowers informaiton. PPdai classified borrowers into seven levels rating, from high to low is AA to HR. According to Fig.1, HR, E and A are the three major credit rating of borrwers, which account almost 89% of total. Fig.2 is the NPL rate of each credit rating. It shows borrowers with HR has the highest NPL rate, which is 45.4%.

7 Figure.1. Credit rating distribution of the sample Figure.2. NPL rate of different credit rating Fig.3. illustrates the distribution of is the repayment period distribution of sample data. According to Fig.3., 6 month and 12 month are the most popular repayment period amount the borrowers in PPDai, almost 80% of them choose 6 and 12 as the repayment period. Fig.4. shows the repayment period is longer, the possibility of NPLs is higher. 18 month is an exception. Figure.3. Repayment period distribution Figure.4. NPL rate of different repayment period Figure.5. NPLs rate of different degrees Figure.6. NPLs rate of different gender Fig.5. is the result of NPLs rate of different degrees. The result indicates borrowers with

8 higher education level have less tendency to be NPLs. Fig.6 shows that women are less likely to break the contract compared with men because the NPLs rate of women is relatively lower. The result is consistent with Schreiner s finding on the credit market (Schreiner, 2004). Considering some variables are continuous and others are discrete, Spearman s correlation analysis is applied to explore the correlationship between the features and loan status. Table 3 shows the Spearman s correlation coefficients for the variables. Results show that there is high correlation between credit rating and loan status (-0.400), times of payments and loan status (-0.400) and repayment period and loan status (0.349). However, marital status, car asset, car loan and identity authentication show no significant correlation with loan status based on the significance test.

9 Table 3. Spearman s correlation coefficient for varibales (N= 9,276)

10 4. Empirical study From the perspective of P2P lenders, the primary concern is whether the money lent out would turn to be NPL or not. A reliable predictive model for NPL possibility based on the published information of a borrower could benefit P2P lenders. In this part, the empirical study has been divided into two parts. Firstly, non-parametric random forest model has been used to explore the key indicator of NPLs. Secondly; a binary logistic regression has been employed to establish the model of possibility of NPLs. Random forests (RFs) model is regarded as an enhanced bagging technique, which is a powerful method for constructing a forest of random decision trees. RFs de-correlate the decision trees in the forest via randomization of split attributes that leads to an improvement over traditionally bagged trees and reduces the variance when averaged over the trees (Breiman, 2001). Some of the variables are continuous and some of them are discrete. RF is an approach which is very suitable for this type of data (Qian, Li, Yu, 2010). Therefore, RF will be applied to explore the factor to NPLs. Random forest is a special type of decision tree. Fig.7 demonstrates the process of random forest. The first step of it is to draw n (D1, D2,, Dn) subsets from the original dataset D randomly through the bootstrap sampling method. The next step is that n DTs are constructed according to the n subsets and n. Finally, each DT casts a unit vote for the most popular class, Then optimal results are determined and classification results obtained. Figure.7. Demonstration of Random Forest Methodology The average accuracy of classification has been measured through ten-fold cross-validation, which can be considered as an evaluation of the performance of RFs. In machine learning,

11 ten-fold validation is widely used to assess a method. It can provide a good trade-off between model over-fitting and under-fitting in general according to Huang, Chen and Wang (2007). The mechanism of it ten-fold validation is to create 10 random splits of data. In this experimental setup, nine folds are used to create a model; one fold is left out and used for testing the model Experiments are executed 10 times over 10 random splits. The whole process has been illustrated in Fig.8. The process is finished in Python The result shows the average accuracy of RFs in the paper is 98%. Figure.8. Demonstration of 10-fold validation Variable selection (VS) is an important and difficult step to identify the best subset of predictors among many variables in a predictive model. Redundant predictors could make noises to a predictive model, waste time and degrees of freedom, and potentially their measurements may not be cost-effective. There are 27 variables as the factors to NPLs theoretically. Lenders would get confused to take all of the variables into consideration when they are making decision. Especially, some factors may be strongly correlated. It is necessary to select several quite predictive factors enable lenders make the judgment as almost the same as involved all factors. Variable importance is a common rule used to VS (Segal and Xiao, 2011; Scornet, 2016). In RF model, there are three methods to measure variable importance: the permutation importance, the z-score and the Gini importance. In the study, permutation importance has been applied to choose some predictive factors. The importance of a variable to predict NPLs can be calculated through the increased error after break the link between the variable and loan status. For example, the predictive accuracy is 98% when all of the 27 variables are involved in RF model. Secondly, break the link between each variable with loan status (the link to other covariates also been broke) randomly and then calculate the corresponding predictive accuracy. In this paper, increase in mean squared error (%IncMSE) has been applied. It calculates the mean squared error when the value of a variable has been permutated in the out-of-bag sample. The higher of the value of %IncMSE is, means the corresponding variable is more predictive to NPL. Table 4 represents the %IncMSE of each variable.

12 Table 4 %IncMSE of variables Variable %IncMSE Variable %IncMSE times of payments working certification 0.73 credit limit income certification 0.70 credit rating credit report 0.69 interest 4.14 company size 0.62 times of success 3.72 working years 0.55 total historical loan amount 2.58 authentication offline 0.4 amount 2.27 car assets 0.37 repayment period 2.17 marital status 0.31 times of application 2.12 house assets 0.27 investment type 1.57 house loan 0.25 age 1.46 car loan 0.20 borrow type 0.97 gender 0.19 income 0.94 identity authentication 0.05 degree 0.81 As shown in Table 4, the top three value are the times of payments, credit limit and credit grading, which means the three are the most predictive factors to NPLs. They should be highly focused when lenders make decision. Moreover, interest, the times of success, the historical loan amount also play an important role to predict NPLs. In order to find out whether eliminating some low importance variable would strongly affect the predictive accuracy or not, the paper conducts a comparison analysis as follow. The second column in Table 4 is the predictive accuracy and recall rate when all of the 27 variables involved in RF. The last column is the result of top 11 important variables which are based on the result of Table 3 involved. According to Table 5, the predictive accuracy of this model based on 11 selected variables is higher than 95% which already can perfectly predict the result. In the following process, only the top 11 variables would be used to generate binary logistic regression model to quantify the possibility of NPLs. Table 5. Comparison results all variable involved the former 11 variable accuracy 98.64% 95.91% recall rate 97.94% 92.75% f % 94.33% Since the major factors has been determined through the above method. The following step is to explore the precise effect of each selected variable on the odds of a non-performing loan. By considering the dependent variable in the case is a binary value, a binary logistic regression model has been applied. Besides, the detailed value of the possibility of non-performing loan of a certain borrow is expected to obtain. In this model, the dependent variable represents the possibility of charged-off loan. In equation (1), z is assumed as a continuous number that was not observed; whose meaning is the probability of a NPL. As a result, a higher value of z indicates a higher default probability. In order to transform this continuous number into a number ranging from 0 to 1, the transformation displayed as

13 follows: 1 p z 1 e (1) According to the transformation, non-linear problem become linear problem. Where p represents the probability of a NPL, when p = 1, it means the loan is an NPL; If p = 0, the loan is a normal loan. Hence, the model can descript as follows: z 0 1x1 2x2 3x3 kxk (2) Where x is the explanatory variable and k is the number of explanatory variable (Hosmer and Lemeshow, 2000). Table 5 displays the result of the above mentioned binary logistic regression model. In order to eliminate the effects of variable s dimension, all of the variables needed in the model are standardized. The binary logistic regression model is first estimated with forward stepwise iterative maximum likelihood method. The analysis is repeated with backward stepwise iterative maximum likelihood method. The final results show no significant difference between the two methods. With the exception of investment type, all of the rest variables in the model have significant impacts on the status of a loan. All of the estimated coefficients in the results are significant at 1% level excluding the times of application significant at 5% level. Table 5. Binary logistic regression result Standardized variable β S.E Wald df Sig. Exp(B) Zamount.818*** Zinterest *** Zmonths 1.899*** Zage.561*** Ztimes of application.211** Zcredit limit -.467*** Ztimes of success 6.537*** Ztotal historical loan.521*** amount Ztimes of payments *** Zcredit rating *** Zinvestment type Constant ** represents significance at the 5% level, and *** represents significance at the 1% level. As shown in Table 5, several implications can be generated: (1) the higher the amount of loan is, the higher risk of NPL will be. (2) The longer the repayment period will cause the higher NPL risk. (3) The older a borrower get, the higher NPL risk will be. Based on the regression results, the probability of NPL for a certain loan is generated with the estimated coefficient as reported in Table 5.

14 z 0.818* zamount ( 1.423)* zinterest 1.899* zrepayment _ period 0.561* zage 0.211* ztimes of application ( 0.467)* zcredit limit 6.537* ztimes of success ( 0.521)* ztotal historical loan amount ( 8.248) * ztimes of payments ( 5.390)* zcredit rating 1 When p z 1 e >0.5, the corresponding loan are more likely to be NPL, and if p < 0.5, the loan are less likely to be NPL. 5. Conclusion and discussion In the information asymmetric market place, loan by mortgaged property could transfer effective information through collaterals. However, P2P loan is a type of loan without collateral and intermediary institution. The only way to decrease the negative impact of information asymmetry is to dig out more useful key points from the available information. This paper has analyzed the determinants of NPL through non-parametric random forest algorithm. A credit risk evaluation model which can be used to quantify the NPL risk of each P2P loan has been established as well. The results show loan with lower credit rating and longer repayment period, are more likely to become NPL, which is consistent with the result obtained by (Emekter et al. 2014). However, gender, education level and marital status of borrowers in the study does not significant effect on NPL. Apart from some demographic characteristics, the historical loan information shows strong connection to NPL. According to the prediction accuracy, the proposed model in the study is robust. The findings of the paper have several implications in terms of selecting the trustworthy borrower of P2P lenders. The historical loan information of a borrower is quite important because they are better reflection of the borrower s credit status. The more successful bids do not mean less likely to be NPL. On the contrary, if a borrower has more successful bids, he or she is more likely turn the loan into bed debt. Although the credit rating offered by the P2P platform is not the value of credit system of China, it is still a valuable indicator of a borrower s credit. Borrowers with high credit limits are turned to be less likely having NPL behavior, it might because they have better financial situation. Since there is no sound supervision mechanism in China to manage P2P lending and disconnection between P2P lending and the official credit system, the lenders can only rely on the published information to decide whether lend the money or not. The paper provides a considerable model for lenders to identify the factors to NPLs and also quantify the possibility of NPLs. According to the result, the RF models show great predictive accuracy and historical loan information of a borrower play a very significant role in predicting NPLs. However, if a borrower has no historical loan information before, the indicator might be different. One of the weaknesses of the paper is that all of the borrowers in the data sample have historical loan information. Hence, the paper does not build a model for those fresh men of P2P lending. In the future study, it is necessary to generate a model by using non-historical loan information data. The combination of the two situations will be more comprehensive. Moreover, the paper just applied the data of one platform; it can use data from several different platforms to see whether the model can be used universally or not. (3)

15 References Cai, S., Lin, X., Xu, D., & Fu, X. (2016). Judging online peer-to-peer lending behavior: a comparison of first-time & repeated borrowing requests. Information & Management, vol. 53, no. 7, pp Diamond, D. W. (1984). Financial intermediation & delegated monitor. Review of Economic Studies, vol. 51, no. 3, pp Emekter, R., Tu, Y., Jirasakuldech, B., & Lu, M. (2015). Evaluating credit risk & loan performance in online peer-to-peer (p2p) lending. Applied Economics, vol.47,no.1,pp Hosmer, D. & Lemeshow, S. (2000) Applied Logistic Regression, 2nd edn, John Wiley, New York. Iyer, R., Khwaja, A. I., Luttmer, E. F. P., & Shue, K. (2009). Screening in new credit markets: can individual lenders infer borrower creditworthiness in peer-to-peer lending? Social Science Electronic Publishing, (rwp09-031). Lee, E., & Lee, B. (2012). Herding behavior in online p2p lending: an empirical investigation. Electronic Commerce Research & Applications, vol. 11, no. 5, pp Lin, M., Prabhala, N. R., & Viswanathan, S. (2013). Judging Borrowers by the Company They Keep: Friendship Networks & Information Asymmetry in Online Peer-to-Peer Lending. INFORMS. Ghosh, A. (2015). Banking-industry specific & regional economic determinants of non-performing loans: evidence from us states. Journal of Financial Stability, vol. 20, pp Gorton G, Winton A (2003) Financial intermediation, Constantinides GM, Harris M, Stulz RM, eds. H&book of the Economics of Finance, vol. 1, part 1, chap. 8 (Elsevier, Amsterdam), pp Huang, C. L., Chen, M. C., & Wang, C. J. (2007). Credit scoring with a data mining approach based on support vector machines. Expert Systems with Applications, vol. 33, pp Klein, N. (2014). Non-performing loans in cesee: determinants & impact on macroeconomic performance. Social Science Electronic Publishing, vol.13, no. 72. Oni, O. A., O. I. Oladele, & I. K. Oyewole Analysis of Factors Influencing Loan Default among Poultry Farmers in Ogun State Nigeria. Journal of Central European Agriculture, vol. 6, no. 4, pp Schreiner, M. (2000). Credit scoring for microfinance: can it work? Journal of Microfinance, vol.2, no.2, pp Schreiner, M. (2004). Scoring arrears at a microlender in bolivia. Journal of Microfinance, vol.6, no.2. Scornet, E. (2016). On the asymptotics of r&om forests. Journal of Multivariate Analysis, vol.

16 146, pp Segal, M., & Xiao, Y. (2011). Multivariate r&om forests. Wiley Interdisciplinary Reviews Data Mining & Knowledge Discovery, vol. 1, no.1, pp Spence, M. (2001). Signaling in retrospect & the informational structure of markets. Nobel Prize in Economics Documents, vol. 92, no.3, pp Tanaka, K., Kinkyo, T., & Hamori, S. (2016). R&om forests-based early warning system for bank failures. Economics Letters, vol.148, pp Tao, Q., Dong, Y., & Lin, Z. (2017). Who can get money? Evidence from the Chinese peer-to-peer lending platform. Information Systems Frontiers, pp Wang, H., Greiner, M., & Aronson, J.E. (2009). People-to-people lending: the emerging e-commerce transformation of a financial market. Value Creation in E-Business Management, vol. 36, no. 13, pp Wu, T. C., & Hsu, M. F. (2012). Credit risk assessment & decision making by a fusion approach. Knowledge-Based Systems, vol. 35, no. 6, pp Lin, X., Li, X., & Zheng, Z. (2016). Evaluating borrower s default risk in peer-to-peer lending: evidence from a lending platform in China. Applied Economics, pp. 1-8.